New CT-based model enhances accuracy in maize endosperm segmentation

The newly proposed CSFTU-Net model integrates advanced attention mechanisms and optimized loss functions to significantly enhance accuracy. This method allows the extraction of key phenotypic parameters such as kernel volume and endosperm composition, offering valuable insights for modern crop improvement.

Maize, a staple crop worldwide, consists mainly of the germ, pericarp, and endosperm, the latter accounting for nearly 90% of kernel dry weight. Within the endosperm, the vitreous region provides strength and resistance to damage, while the starchy region is softer but more vulnerable to pests and diseases. Accurate quantification of these two regions is vital for understanding kernel texture, grain quality, and suitability for processing. Conventional imaging methods, however, often fail to distinguish between regions due to blurred boundaries and low contrast, limiting their usefulness in breeding programs. Based on these challenges, a nondestructive, high-resolution imaging and analysis approach is needed to advance maize phenotyping.

A study (DOI: 10.1016/j.plaphe.2025.100022) published in Plant Phenomics on 28 February 2025 by Xinyu Guo’s ＆ Chunjiang Zhao’s team, Beijing Academy of Agriculture and Forestry Sciences & Shanghai Ocean University, provides plant scientists and breeders with a powerful tool to evaluate maize kernel structure rapidly and precisely.

The study introduces a deep learning–based CT image analysis framework to overcome limitations of traditional segmentation methods. Kernels were batch-scanned using a multislice CT scanner, and individual kernels were extracted using image processing algorithms. The core innovation, the CSFTU-Net model, improves on the widely used U-Net architecture by incorporating convolutional block attention modules (CBAM) in the encoder and squeeze-and-excitation (SE) mechanisms in the decoder, alongside a focal Tversky loss function and boundary smoothing. These modifications boost the model’s sensitivity to fine structural details and blurred edges between vitreous and starchy endosperm. When tested on 1,000 CT datasets, CSFTU-Net achieved a Dice coefficient of 89.13%, outperforming other state-of-the-art models such as DeepLabv3, TransU-Net, and Swin UNETR. This high accuracy enables reliable extraction of phenotypic parameters, including kernel volume, vitreous endosperm volume, starchy endosperm volume, and their respective ratios. A dataset of 250 maize varieties further demonstrated the model’s ability to identify subgroup differences, highlighting varietal variation in endosperm composition. These quantitative insights are directly relevant for breeders aiming to optimize varieties for specific industrial applications.  For instance, vitreous-rich kernels were linked to food products requiring hardness, such as cornflakes, while starchy-rich kernels suited milling and oil extraction.

By quantifying the proportion of vitreous versus starchy endosperm, the system supports targeted breeding for traits such as grain hardness, nutritional quality, and mechanical resilience. For food industries, the findings offer practical guidance: vitreous endosperm, rich in amylose, is better suited for cornflakes and puffed products, while starchy endosperm, rich in amylopectin, is preferable for cornmeal and oil production. In addition, the framework sets a precedent for extending AI-driven CT analysis to other cereals, paving the way for broader applications in smart agriculture and crop processing industries.

###

References

DOI

10.1016/j.plaphe.2025.100022

Original URL

https://doi.org/10.1016/j.plaphe.2025.100022

Funding information

This research was funded by the National Center of Pratacultural Technology Innovation (under preparation) Special fund for innovation platform construction (CCPTZX2023K03),Industrial Technology Innovation Program of IMAST(No. 2024RCYJ04004) and the National Natural Science Foundation of China (32271987).

About Plant Phenomics

Science Partner Journal Plant Phenomics is an online-only Open Access journal published in affiliation with the State Key Laboratory of Crop Genetics & Germplasm Enhancement, Nanjing Agricultural University (NAU) and distributed by the American Association for the Advancement of Science (AAAS). Like all partners participating in the Science Partner Journal program, Plant Phenomics is editorially independent from the Science family of journals. Editorial decisions and scientific activities pursued by the journal's Editorial Board are made independently, based on scientific merit and adhering to the highest standards for accurate and ethical promotion of science. These decisions and activities are in no way influenced by the financial support of NAU, NAU administration, or any other institutions and sponsors. The Editorial Board is solely responsible for all content published in the journal. To learn more about the Science Partner Journal program, visit the SPJ program homepage.